Will AI Progress Accelerate or Slow Down? Projecting METR Time Horizons

A new analysis from METR (formerly ARC) examines whether AI progress will accelerate or slow, using time horizon—the length of tasks AI can complete—as a key metric. Their data shows the 50% time horizon (tasks AI completes with 50% success) has been doubling every 196 days historically, but since 2024, that rate has accelerated to just 89 days. This suggests we may be entering a period of dramatically faster capability growth, with the potential for AI to reach "infinite time horizons" where models can handle arbitrarily long, complex tasks.

The report outlines several AI-driven feedback loops that could accelerate progress further: data generation (AI creating synthetic training data), coding (automating AI R&D tasks), research taste (AI setting research directions), chip technology (AI designing better hardware), chip production (automating manufacturing), and economic feedback (AI boosting overall investment). These loops create self-reinforcing cycles where smarter AIs accelerate the development of even smarter systems.

Forecasters are using METR's time horizon metric in initiatives like the AI 2027 scenario and AI Futures Model because it has no performance ceiling and correlates strongly with other capability measurements. The analysis presents multiple projection scenarios: continued 89-day doubling (segmented exponential), reversion to 196-day doubling, or smooth superexponential growth as feedback loops intensify. Distinguishing between these scenarios will be crucial for predicting when transformative AI capabilities might emerge.